Reducing the intrusive driving behaviour in lane departure avoidance system using machine learning approach

This paper proposes two components to reduce the intrusive behaviour in Lane Departure Avoidance System (LDAS) during safety intervention. The components are; i) human mimic Driver Model (DM) for lane recovery and keeping assessment, and ii) Predictive Model (PM) that has the ability to predict unintentional lane departure. The lane departure prediction is used to minimise the lane keeping steering angle (μk) output from the Lane Keeping Driver Model (LKDM) and the lane recovery steering angle (μr) from the Lane Recovery Driver Model (LRDM). With this design, a small corrective input (steering angle) that guides the vehicle to compensate disturbance (unintentional lane departure) shall reduce the intrusive behaviour during safety intervention of the LDAS. Both model, LKDM and LRDM, are combined to form a single DM. Subsequently, both of the DM and PM are merged to form the LDAS. Two methods were used to formulate the LDAS; i) Neural Network (NN, benchmark approach) and ii) Nonlinear Autoregressive with Exogenous Inputs (NARX, proposed solution). Both LDAS models are evaluated against three road curvatures (ρ = 800m, 3000m, 10000m) and injected with a simulated unintentional lane departure disturbance. The response of each LDAS model is analyzed against the intrusive driving assessment data. The result shows that with a sufficient prediction of time to lane crossing (tLC) provided by the PM, the DM is able to provide a small lane recovery and lane keeping steering angle thus ensuring the LDAS to operate within the non-intrusive driving behaviour. This study is important in providing a more human-behavior like LDAS performance.